This thesis presented a method for estimating the position of a mobile robot, without an a priori estimate. This is accomplished by learning a set of visual features, known as landmarks, candidates for which are detected as local maxima of a measure of distinctiveness. Specifically, edge density is employed as the measure of distinctiveness. Landmark candidates are then grouped into tracked landmarks: sets of candidates which correspond to the same visual region of the environment, as observed from different viewpoints. Grouping is achieved by matching subspace encodings of the candidates, perhaps with adjustments in position in the image in order to improve matching. Online position estimation is performed by detecting candidates and matching them to the tracked landmarks. Each match is used to generate a pose estimate by employing a principal components reconstruction of a feature vector which encodes both appearance and image geometry. The experimental results indicate that the method is robust for a variety of environments and parameterisations and shows promise for a range of applications.